![\"Internal](\"https:\/\/storage.googleapis.com\/arrowlancer-prod.firebasestorage.app\/blog-ma\/assistant-ia-interne-maroc-2026\/1.jpg\")
<\/p>\nAn internal AI assistant \u2014 a chatbot plugged into your own documents \u2014 costs between 80,000 MAD for a serious pilot and 600,000 MAD for a multi-source production platform in Morocco in 2026. A realistic timeline runs from six weeks to six months. And a large share of these projects die somewhere between the demo and go-live.<\/p>\n
It is almost never the model\u2019s fault. Today\u2019s models are more than good enough to answer an HR question or find a clause in a contract. What breaks a project is the real state of your documents, a scope sold too broadly too fast, and teams nobody bothered to convince that they should change how they work.<\/p>\n
This article is written for exactly where you are: the demo impressed you, you are about to open a vendor consultation, and you want reference points before you receive quotes that differ by a factor of six.<\/p>\n
An AI assistant demo takes two days to prepare. The vendor picks ten to twenty clean, well-structured documents in a single language, indexes them, and asks three questions whose answers they already know. The result is impressive. It is also a poor picture of your company.<\/p>\n
Your reality looks more like this: 8,000 files scattered across a file server, a Google Drive, three people\u2019s mailboxes and an ERP; procedures in French, customer threads in transliterated Darija, contracts in Arabic; four versions of the same document with no agreement on which one is authoritative; and PDFs scanned crooked back in 2019 that nobody has reread since.<\/p>\n
The gap between the demo and that reality is not a matter of polish. It is where most of the budget and most of the risk live. An honest provider tells you this before you sign; a salesperson lets you discover it in sprint three, once the money is committed.<\/p>\n
The word \u201cassistant\u201d hides three families of use cases with very different cost and difficulty. Mixing them in a single RFP is the number one reason your quotes will diverge: two honest vendors can read the same request and propose 120,000 and 450,000 MAD, simply because they understood different things.<\/p>\n
<\/p>\n
RAG means retrieval-augmented generation. In practice: your documents are split into chunks, turned into vectors and stored in a database; for every question the system retrieves the five to ten relevant passages and asks the model to write an answer from those passages only, with sources cited. The model does not \u201clearn\u201d your data \u2014 it reads it on demand. This is the foundation of nearly every internal assistant, and the easiest part to build.<\/p>\n
Same engine, but exposed to users who are not forgiving: your customers, or your own staff in a hurry. Now you must handle the cases where it does not know, handover to a human, conversation history, tone, and guardrails on what it must never say (commitments on price, delivery dates, warranty terms). Expect 40 to 60% more work than an internal-only RAG.<\/p>\n
Here the assistant never speaks to the customer \u2014 it prompts the agent. It drafts a reply, pulls up history, summarises a case file. For most Moroccan SMEs this is the smartest starting point: visible error risk is low, a human stays in the loop, and the time saved is measurable in the first month. Many companies should start here rather than with a public-facing chatbot.<\/p>\n
These figures apply to companies of 20 to 200 employees working with a Morocco-based provider. They move with the number of sources<\/em> to connect and the quality bar you set, far less with the number of pages.<\/p>\n Then there are the recurring costs most quotes forget to price:<\/p>\n One line deserves to be isolated in the quote: data preparation<\/strong>, typically 15 to 30% of the total. If it appears nowhere, it has been underestimated \u2014 or it will reach you as a change order. For the wider budget picture across other AI project types, our AI project cost guide for Morocco<\/a> breaks down the remaining cost lines.<\/p>\n The timeline quoted in pre-sales is almost always the technical timeline. The real one includes your own delays: finding who owns which document, getting access granted, having business teams review answers.<\/p>\n In short: 6 to 8 weeks<\/strong> for a genuinely narrow pilot, 3 to 6 months<\/strong> for an assistant your teams open every day. Any promise of \u201cdelivery in three weeks\u201d on a broad scope is describing a demo, not a product.<\/p>\n An assistant cannot be better than your documents. If three versions of the returns procedure exist, it will confidently cite the wrong one. Before indexing anything, you need one authoritative repository, dated documents, archives excluded, and a clear decision on who may see what. That is precisely the work described in our piece on the data quality audit before a BI or AI project<\/a> \u2014 and it belongs before the AI vendor, not during.<\/p>\n A good RAG assistant cites its sources and can say it does not know. The real problem is not that it is sometimes wrong \u2014 it is that nobody on the project measures how often. From day one, require an evaluation set of 100 to 200 real questions with expected answers<\/strong>, written by your business teams, and a correctness figure at every delivery. On a well-bounded scope, 80 to 90% correct is a good result; nobody can seriously promise 100%, and a vendor who claims it is not measuring anything.<\/p>\n The classic failure: the assistant works, ten people try it in week one, two still use it a month later. The causes are always the same \u2014 the assistant lives in a tab nobody opens instead of inside the tool people already use; the first mediocre answers were never corrected; and no manager ever said out loud that this is now the default reflex. Plan for an internal champion, a one-click way to flag bad answers, and a monthly review of unanswered questions. That review is what actually makes the tool improve.<\/p>\n Microsoft 365 Copilot, ChatGPT Enterprise and their equivalents are priced per user per month (a few hundred MAD per person \u2014 check the current figure with the vendor, it changes fast). For 40 users that often lands at 120,000 to 200,000 MAD per year with zero development. That is not trivial next to a custom project amortised over three years.<\/p>\n The practical rule: if your documents already live in the vendor\u2019s ecosystem<\/strong> (SharePoint, Drive, mail) and your needs are generic \u2014 summarise, draft, find a file \u2014 SaaS almost always wins. If the value sits in a business repository<\/strong> (technical catalogue, contracts, ticket history, regulatory library) or in a homegrown ERP, custom becomes the sensible option, because no subscription will read those sources properly. The full reasoning is in our comparison of SaaS versus custom software<\/a>.<\/p>\n A third criterion carries real weight in Morocco: personal data protection. As soon as the assistant touches employee files, customer records or ID documents, you are within the scope of Law 09-08, with its questions of purpose, retention and transfer outside Morocco. We covered this in depth in our article on CNDP compliance for ChatGPT and Copilot at work<\/a>; for your specific case, confirm the applicable formalities on the CNDP website before you settle the hosting question. It is also one of the trade-offs we work through in our AI and digital transformation engagements<\/a>.<\/p>\n Take this list to your next meeting. The answers will tell you more than a second demo.<\/p>\n A solid provider answers questions 2, 3 and 12 straight, even when the answer is uncomfortable. Evasiveness on those three is the clearest warning signal available to you before you sign.<\/p>\n Arrowlancer designs and ships AI solutions for Moroccan companies: internal assistants and document RAG, automation, document analysis, and sovereign AI hosted in Morocco when your data requires it. We usually start with a scoped, measured pilot rather than a demo.<\/p>\n\n
\n
How long it takes, from scoping to go-live<\/h2>\n
\n
What derails the project: data, hallucinations, adoption<\/h2>\n
Data nobody has actually looked at<\/h3>\n
Hallucinations, and the missing measurement<\/h3>\n
Adoption, which is decided before the tech<\/h3>\n
Generic AI SaaS or a custom assistant: how to decide<\/h2>\n
Questions to ask a vendor before signing<\/h2>\n
\n
What now?<\/h2>\n